name: qiaomu-codex-imagegen description: | Generate or edit images with Codex's built-in image generation from any agent, through an MCP tool or a plain CLI. For every image request it first proposes at least four divergent style directions (24 distilled templates, 48 presets, 20 Mondo poster artists), lets the user pick, expands the prompt, generates, and checks the result. Scenario craft for 视频封面, 视频海报, 小红书配图, 公众号封面, X 封面, 朋友圈海报, 书籍封面, 专辑封面, 产品海报, 人物写真, 节气海报, 展览海报, PPT 页面, 图生图. Use for 生图, 配图, 海报, 封面, "给我几个风格", "用 Codex 画一张". Excludes: editing photos with local tools, charts or diagrams from data, UI mockups, video generation, or anything that needs an API key-based image service. license: MIT metadata: version: 0.3.0 author: 向阳乔木 (https://x.com/vista8)

Qiaomu Codex ImageGen

Codex ships an image generator that other agents cannot call. This package relays to it (codex app-server) and adds the craft that makes the pictures usable: a library of 24 templates (48 presets) distilled from 689 reference prompts, Mondo poster techniques, scenario presets, and a rule set for facts, text and identity.

How to call it

1. MCP tools present (suggest_directions, compose_prompt, generate_image, search_prompts, get_prompt, build_prompt, list_catalog; in Claude Code mcp__qiaomu-codex-imagegen__*): use them.
2. No MCP: run the CLI from this skill's folder: node scripts/cli.mjs suggest|compose|generate|search|prompt|list (see --help).
3. Neither available: tell the user to install (README). Requires the Codex CLI, logged in. Never fake an image with code.

The workflow: suggest, choose, compose, generate, check

1. Intake (no tool call yet)

Extract: topic, deliverable (what it is for), ratio, text mode, facts that must be exact, references and the role of each (identity person, product real packaging, style, layout), must-keep, must-avoid.

• Text mode: none (no text), exact_short (only the short text you list), typeset_later (text-free base with room for the title).
• Posters, covers, cards and slides need a typographic layer. A finished poster has a hierarchy: headline, subtitle, one or two info lines, a small tag. With only a headline, or none, the result is concept art, not a poster. So for these deliverables use exact_short and give a complete short copy set: headline plus copy as lines separated by ｜. Take the lines from the user's facts. If they gave none, ask, or write clearly fictional sample lines and say so; never invent real prices, dates of real events, names, statistics.
• none fits portraits and photos (T09–T11), product macros that carry no information, and base images the user will typeset. Keep every string short and read the result: models misspell, especially long Chinese.
• Ask only blocking questions, at most three: a missing subject, a real person's identity reference, or facts that must appear. Fill low-risk gaps with a stated default.

2. Suggest at least four divergent directions

Call suggest_directions with the topic, deliverable and any headline/text mode. It returns four or more directions with different visual mechanisms (typography-led, graphic structure, photographic, product macro, material/space, plus a Mondo artist option for poster-like work). Never mix two mechanisms into one image.

Present them to the user briefly, one numbered item each:

• name and template id,
• one concrete sentence of what this topic would look like in that mechanism (write it yourself from mechanism and the topic, not a copy of the lock text),
• what you still need from them (subject, data, reference photo),
• look at one references[].image thumbnail per direction when available, so your sentence matches the real look.

Then ask which to make: one, several, "all", or "你定". Offer to exclude these and propose more (exclude).

Skip the question when the user already fixed the style ("用 Mondo 风格", "T03", "就这个风格") or said to go ahead ("直接出", "你定"). For "你定", pick the two best-fitting directions from different families and say which and why.

3. Compose the prompt

For each chosen template direction call compose_prompt with concrete variables:

• Map the topic to one visible silhouette, action or relation, and say how it connects to the topic. At most one main symbol and one cross-over action.
• Colours are roles (background, primary, accent, ink), not fixed hex values. Material belongs where the preset puts it; do not overlay noise on everything.
• No invented facts: prices, statistics, titles, names, dates, weather, quotes, QR codes, logos.
• missing_required non-empty → ask for exactly that. weak_variables → they still hold generic defaults, fill them first.
• Preset with structural_override → rewrite the prompt by hand: delete the base sentences the override replaces, keep no conflicting structure, then send the result as raw_prompt with the 避免：… line.
• Typographic templates with no text: tell the user the base image only satisfies the style after the title is typeset.
• Give copy as ｜-separated lines; when strings include dates or numbers, list them all so the whitelist matches exactly (or pass a full text_rule).

Mondo directions skip this step: generate_image { prompt, style: <artist>, preset: <video-cover|poster|…> }.

Full rules and the canonical agent instruction: references/design-system/agent-guide.md.

4. Generate

Default one image per chosen direction, up to four, each its own generate_image call (they can run in parallel). Use count only for variants of the same direction. Say that it takes 30–180 s and spends Codex quota. out_dir goes inside the user's project when they name one; reference_images take absolute paths (identity photo, product shot, style reference, or the image to edit).

Series: lock template, type hierarchy, colour roles, texture placement and cross-over action; change one or two of scene, narrative focus, hue, module span. Pass the first image as reference_images for the rest.

5. Check and deliver

Read each saved image and test it against acceptance_checks (the tool prints them): first-glance focus, relations that must hold, materials only where intended, exact text and identity, nothing invented. State pass or the specific failed relation.

With the local corpus pack installed, compare against the references: put the result next to one or two references[].image of the same template (suggest_directions lists them). If yours is plainly less designed (bare subject, no typographic hierarchy, no secondary layer), that is a failure too: complete the copy set or the missing layer and regenerate once.

On failure change only the sentence that carries that relation and regenerate once; do not stack "more premium" adjectives. Transparent regions in a result are flattened onto white automatically (transparent_background: true keeps them). Deliver absolute paths, the direction used, and the sidecar .json (exact prompt) next to each image. Only say "generated" for images you actually produced and looked at.

Rules

• Each generation spends the user's quota: no test generations, count above 3 only if asked.
• People: prefer silhouettes, back views, hands or symbols unless a face is asked for. Identity consistency needs an identity reference photo; without one say the person is fictional. Never make deceptive material with a real person's likeness.
• Facts and numbers come only from the user; complex charts and long copy are typeset afterwards, not drawn by the model.
• Reference corpus: search_prompts / get_prompt read the local pack of 689 reference prompts (installed on the user's machine only). Use them to study how a mechanism is described; never send a corpus prompt unchanged as the user's prompt.
• Simple route: prompt + preset + style (66 styles, 20 Mondo artists) when the user wants a quick image without the direction step.

Output contract

Reply with: absolute path(s), pixel size, direction (template/preset or Mondo artist) per image, assumptions made, acceptance result, and where the prompt sidecar is. On failure give the tool's error and the one fix to try (Codex login, shorten the prompt, check the reference path).

Resources

• Agent guide: the rule set, input/output shapes, routing table
• Template library: 24 templates, locks, presets, evidence; machine-readable templates.json, routing.json; filled examples
• Scenarios: ratios and safe areas per platform; prompt craft; Mondo poster craft with artist styles
• references/styles.json (66 styles), mondo-artists.json (20), presets.json (13 scenarios)

Copyright (c) 向阳乔木. X https://x.com/vista8, GitHub https://github.com/joeseesun/. Template library distilled from public-area samples of https://vip.xiaoxiaodong.ai/open-source; Mondo material from qiaomu-mondo-poster-design (MIT).

乔木 Codex 生图 · qiaomu-codex-imagegen

让任何 Agent 都能调用 Codex 内置的生图能力：每次生图先给出至少四种机制不同的风格方向，你选定后再出图。 内置 24 类场景模板（48 个预设）、20 位 Mondo 海报设计师风格，以及一套防止编造事实、乱写文字的规则。

Codex 的生图工具只在 Codex 里能用。这个项目把它接出来：一个 MCP 服务、一个命令行、一个 Agent Skill，共用同一套核心。Claude Code、Codex、Cursor，或者任何能执行命令的 Agent，都能说一句"做一张小红书配图"就拿到图片文件。

Use Codex's built-in image generation from any agent through MCP, a CLI or a skill. Every request first gets at least four divergent style directions (24 templates, 48 presets, 20 Mondo artists); you pick, it composes the prompt, generates, and checks the result.

先给方案，再出图

对 Agent 说「给这期视频做一张封面」，它不会直接画，而是按下面的顺序走：

1. 推荐方向：suggest_directions 给出 至少四个机制各异的方向，来自不同家族（文字主导、平面结构、摄影人物、产品近摄、材料空间，再加一个 Mondo 设计师方向），每个方向带风格锁、需要你补的信息和参考案例缩略图。
2. 你来选：选一个、几个，或者说「你定」。想要更多就排除这几个再推荐。
3. 展开提示词：compose_prompt 把变量填成完整提示词，同时给出避免项、验收条件、假设和缺失信息。带「结构替换」的预设会提示 Agent 改写，不会把冲突的句子硬拼在一起。
4. 生成：每个选中的方向出一张，图旁边留一份 JSON，记录实际提示词，方便复现和做系列。
5. 验收：Agent 看图，对着验收条件逐条检查；不合格就只改对应那一句再生成一次。

你已经指定了风格（「用 Mondo 风格」「T03」）或说「直接出」时，会跳过选择这一步。

能做什么

样例

下面所有图片都是用这个工具调用 Codex 生成的，没有后期修图；活动、品牌、人物和文案全部是虚构的示例。每张图的提示词和参数在 docs/samples/prompts.json。

同一个主题，五个方向

主题都是「咖啡店阅读月」。suggest_directions 给出的方向机制完全不同，所以出来的不是同一张图换个滤镜：

24 个类别各一张

场景样例

改图

这些样例是怎么做出来的

• 每个类别都走完整流程：推荐方向、填变量、生成、对着参考案例验收。
• 第一版不合格，重做了：最初我只给了标题甚至不给文字，出来的是干净的概念图，不是海报；并排对照参考案例后，补全了副标题、信息行和角标这一层文字，才有现在的完成度。肖像（T09–T11）按方法论保持无字。
• 第一版还暴露了一个工具缺陷：约四分之一的图带透明通道，在看图软件里显示成黑边。现在提示词会要求背景不透明，保存后检测到透明区域会自动压平到白底（transparent_background: true 可保留）。
• 图内文字：标题和主要文案已逐张核对；小字号的信息行仍可能有细微瑕疵，正式使用请放大检查。
• 肖像是 AI 生成的虚构人物，不对应真实个人。

安装

前置条件：

• Node.js 18+（node --version）
• 已安装并登录的 Codex CLI（codex --version 能运行）
• Codex 账号可使用图片生成（先在 Codex 里手动生成一张试试）

1. 作为 Skill 安装（任何支持 Agent Skills 的工具）

npx skills add joeseesun/qiaomu-codex-imagegen

Skill 会让 Agent 知道怎么选场景、风格、怎么写描述，并通过下面的 MCP 或自带 CLI 出图。

2. 注册 MCP（推荐，Agent 直接调用工具）

Claude Code：

claude mcp add --scope user qiaomu-codex-imagegen -- node /绝对路径/qiaomu-codex-imagegen/scripts/mcp-server.mjs

其他 MCP 客户端（Cursor、Codex 等），在配置里加：

{
  "mcpServers": {
    "qiaomu-codex-imagegen": {
      "command": "node",
      "args": ["/绝对路径/qiaomu-codex-imagegen/scripts/mcp-server.mjs"]
    }
  }
}

提供七个工具：

3. 只用命令行

git clone https://github.com/joeseesun/qiaomu-codex-imagegen && cd qiaomu-codex-imagegen
node scripts/cli.mjs "窗边翻开的书，书页上升起咖啡的热气" --preset xiaohongshu

你可以直接这样说

装好 Skill 和 MCP 后，对 Agent 说：

• 「给这期视频做一张 16:9 的封面，要有设计感，先出三个方案」
• 「做一组小红书配图，主题是清晨读书，第一张定风格，后面几张保持一致」
• 「用 Olly Moss 风格做《沙丘》的电影海报，不要放字」
• 「把这张图的背景换成米色纸张，主体不变」

用法示例（命令行）

# 先要方向（免费），再选
node scripts/cli.mjs suggest "一个月的咖啡店阅读活动" --for 视频封面

# 展开模板，看完整提示词和验收条件（免费）
node scripts/cli.mjs compose --template T03-1 --var topic=谷雨 --var subject=嫩芽 --var headline=谷雨

# 带模板生成
node scripts/cli.mjs generate --template T03-1 --var topic=谷雨 --var subject=嫩芽 --var headline=谷雨

# 搜本地案例
node scripts/cli.mjs search "节气 茶" --limit 5

简易路线（不走方向推荐）：

# 视频封面，三个方案先看
node scripts/cli.mjs "一个巨大的红色播放键像日出一样升起" --preset video-cover --style saul-bass --count 3

# Mondo 风格电影海报，只出图不放字
node scripts/cli.mjs "沙漠中巨大的沙虫轮廓，渺小的人影" --preset movie-poster --style olly-moss

# 必须带标题时，逐字给出
node scripts/cli.mjs "一座灯塔" --preset wechat-cover --text "夜航船"

# 图生图
node scripts/cli.mjs "保持同一支铅笔，背景换成暖米色纸张" --ref /abs/pencil.png --ratio 1:1

# 先看提示词，不生成
node scripts/cli.mjs "咖啡和书" --preset xiaohongshu --show-prompt

参数：--preset 场景，--style 风格（键或自由文字），--ratio 覆盖比例，--text 图内文字（可重复），--ref 参考图（可重复，绝对路径），--count 并行变体 1–4，--out 输出目录，--name 文件名，--model、--timeout，--json 机器可读输出。默认保存到 `~/Pictures/qiaomu-codex-ima